Start with what you are getting
An open-weight model makes its trained weights available for others to download, subject to its license. You can usually run it using compatible software on your own hardware or rented computing resources. That does not necessarily mean the training data, training code or full development process is public. It also does not mean every use is permitted: licenses can impose conditions or restrictions.
A hosted AI API gives you access to a model through a provider’s service. Your application sends a request and receives a response without operating the model’s infrastructure. The provider typically manages the servers, model deployment and much of the surrounding software. Some providers host open-weight models too, so the distinction is not simply which model you choose. It is also who operates it and which parts of the system you can change.
Control over behavior has limits on both sides
With open weights, you can choose the model version, set up your own prompts and supporting tools, and often adjust how responses are generated. Depending on the model and your resources, you may also fine-tune it for a particular task. You can keep an older version running while you test a newer one. These choices are useful when a workflow needs predictable changes or specialized output.
None of that guarantees a particular answer. Model behavior still depends on its training, your input, the software used to run it and any adaptations you make. Fine-tuning can improve a narrow task while making other behavior worse, so it needs evaluation. With a hosted API, you can generally control inputs and exposed settings, but not the underlying weights or unannounced changes to the service. Providers may offer version selection or customization features; the available controls vary by product.
Control over data depends on the whole setup
Running an open-weight model locally can keep prompts and outputs on a device you manage. That can be valuable when you do not want to send working documents to a third-party model service. But local operation is not automatically private: connected apps, logging tools, backups or external search features may still transmit data. If you run the model on rented infrastructure, that provider also becomes part of your data-handling picture.
With a hosted API, your request must reach the provider for processing. Before using it with sensitive material, check the relevant service terms and settings, including retention, logging and any options concerning use of submitted data. Then examine your own application: careless logs or broad employee access can expose information regardless of which model it calls. A practical rule is to send only the data a task needs, and remove identifying details when they are unnecessary.
Operations are a form of control—and responsibility
Self-hosting lets you decide when to update a model, how much computing capacity to provide and what happens if your usual model is unavailable. In exchange, someone must manage deployment, security, monitoring, software compatibility and failures. A model that works on a laptop for occasional questions may require a different setup to serve several people at once. Hardware limits can also constrain model size, response speed or the amount of text processed in one request.
A hosted API shifts much of that operational work to the provider, which can make a first project simpler. You still need to handle errors, usage limits, changing prices and possible interruptions. Costs are not directly comparable from a headline price alone: self-hosting involves hardware or rental charges and maintenance time, while an API commonly charges according to usage. Estimate costs using your expected volume, then test with a small workload before committing.
Choose for the workflow, not the label
For an individual experimenting with prompts or a small team building a prototype, a hosted API may be the easier starting point. For a team that needs to keep a specific version available, adapt a model deeply or run without sending prompts to a model provider, open weights may be worth the operational effort. A hybrid approach is possible: use a local model for sensitive drafts and a hosted service for tasks where its convenience or capabilities matter more.
Test either option on examples from your real work. Check whether the answers are useful, whether errors are easy to detect, how the system handles sensitive input and how it behaves when a request fails. The meaningful question is not which approach offers “full control.” Neither does. It is which controls your workflow needs, which responsibilities you can manage and which limitations you are prepared to accept.
